Skip to content
Genitive.AI
Perspective · September 3, 2026 · By Vivek Hans, Genitive.AI

AI Readiness Is Not Agentic Readiness

An organization can be AI-ready and agentically unready. What changes when software acts, and the five questions an AI assessment never asked.

An organization can be genuinely AI-ready and agentically unready. That is not a failing. It is a different question, and most assessments have not asked it yet.

Most organizations now have some form of AI assessment behind them: a maturity model, a data audit, a governance review, a platform vendor’s readiness questionnaire. The questions were sensible. Can we build and deploy models? Is the data accurate and permitted for use? Is the output good enough to act on? Who reviews it? Then the agents arrived, and the same assessment was reused to decide whether the organization was ready for them.

It was not designed for that. AI produces outputs that a person acts on. Agents take actions that a person may never see. That one shift changes what has to be true before anything scales, and it is why an organization can pass an AI readiness review and still be nowhere near ready to let software act on its behalf.

The gap, in the numbers

Leaders expect agents to remake how work gets done. In Deloitte’s August 2026 survey of 501 senior leaders across five US industries, all at organizations already piloting agentic AI, 74 percent expect nearly half of their business processes to be redesigned or rebuilt around AI agents within four years. In the same survey, 5 percent say their processes are highly prepared for agents, and one in five say their organization is prepared to redesign processes for autonomous operation. These are organizations that have already started, and the readiness figure is still single digits.

The foundations picture matches. McKinsey found in April 2026 that nearly two-thirds of enterprises have experimented with agents, fewer than 10 percent have scaled them to tangible value, and eight in ten cite data limitations as the roadblock. Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, through escalating costs, unclear business value or inadequate risk controls.

Read those together and the failures cluster in a specific place: after the model works, and after the AI assessment stopped asking questions. Cost that only appears at production volume. Risk controls built for outputs rather than actions. Processes that were never redesigned for something that acts inside them.

Two assessments, two questions

An AI assessment asks four things. Can we build, deploy and govern models? Is the data accurate, available and permitted for use? Is the output good enough to act on? Who reviews it before it is used?

Agent readiness asks five different ones. What identity does the agent act under, and whose permissions does it carry? Does the chain hold, or does reliability multiply down across the steps? Who sees what it did, and when? What does it cost per run at full audience, not at pilot? What happens when the model beneath it changes?

The first set is necessary. It is not sufficient, because everything in the second set only starts to matter once something acts.

What changes when something acts

Output becomes action. A recommendation needs a reader. An action needs an identity, permissions scoped to the task, and a named human accountable for what it does. The clearest illustration so far is regulatory. In April 2026 the FDA issued what is being described as its first warning letter relating to AI, to a cosmetics and drug manufacturer that had used AI agents to create product specifications, procedures and master production records. The firm’s response was that the agent had never indicated validation was required. The FDA’s position was that AI-generated output must be reviewed and approved by an authorized representative of the quality unit before it is used. Nobody had asked who was accountable for an output that had already become a record.

One output becomes a sequence. A single model call has a quality score. An agent that takes eight steps has eight of them, and they multiply rather than average. Five steps at 99 percent each land at roughly 95 percent; ten steps at 99 percent land near 90 percent. An assessment that scored the model at 99 percent told you nothing about the workflow.

Known steps become an unknown path. Automation follows a script you can test end to end. An agent chooses its route toward a goal, so the path through the systems is not known in advance. You cannot validate the path. You validate the guardrails, the boundaries the agent is not allowed to cross, and you test that they hold under pressure.

They receive it becomes they may not see it. When AI produces a draft, a person reads it before it goes anywhere. When an agent places an order, adjusts a forecast or emails a customer, the person may find out afterwards, or never. Oversight has to be designed into the workflow rather than assumed from the fact that humans are nearby, and it decays: the better the agent performs, the less anyone looks.

Model calls become ongoing operation. A model in a chat window costs what it is used. An agent running across every transaction costs what the business does, and multi-step workloads are getting more expensive, not less. Gartner projects that inference costs per agentic workflow will rise more than fivefold through 2028 as products shift from single responses to multistep autonomous execution. Behavior also drifts on its own, and when the vendor swaps the model underneath, the agent you validated is no longer the agent you are running.

This is not an argument against AI assessments

The AI assessment is the base, and an organization that fails it should not be building agents at all. Data that is accurate, available and licensed for use, outputs that are evaluated, and a review step that works are the conditions for everything above. In our experience, less than half of what an AI assessment checks carries over to the agentic question; the rest is everything that only starts to matter once something acts. That estimate is ours, not a survey result, and the boundary is fuzzy. A copilot that drafts an email for someone else to send sits between the two: the output becomes an action the moment the person stops reading before clicking send. That is why every candidate workflow should be screened rather than assumed, and why the first screening question is not “how autonomous should this be” but “does it need to act at all”. Many do not. Rules-based automation and better analytics clear the bar for a large share of the workflows currently being pitched as agents, at lower cost and lower risk.

A screen you can run this week

Take one candidate workflow and answer the questions in the right-hand column. If any answer is “nobody knows”, the workflow is an AI use case for now, whatever the vendor calls it.

An AI assessment asksAgent readiness asks
Can we build, deploy and govern models?What identity does it act under, and who is accountable for what it does?
Is the data accurate, available and permitted for use?Does reliability hold across every step, and how is that measured?
Is the output good enough to act on?Which guardrails bound the path it may take, and have they been tested?
Who reviews the output before it is used?Who sees what it did, when, and could they reconstruct why?
Has the value been measured?What does it cost per run at full volume, and what happens when the model changes?

The left column is close to a solved problem in organizations that have taken AI seriously. The right column is where the 40 percent cancellation rate lives.

The question worth asking first

AI readiness is not agentic readiness. The organizations that scale agents will not be the ones with the best AI assessment. They will be the ones that asked the second question before the first agent went live, screened each workflow against it, and let the answer decide what was allowed to act. That is the question our Agentic Readiness Screen exists to answer, one workflow at a time. Aim high. Gate hard.

Frequently asked questions

What is the difference between AI readiness and agentic readiness? AI readiness asks whether an organization can build, deploy and govern models whose outputs a person then acts on: data, model quality, output review. Agentic readiness asks what has to be true when software acts on its own: identity and accountability, reliability across a sequence of steps, tested guardrails, oversight that is designed in, and the cost and drift of continuous operation.

Can an organization be AI-ready but not ready for agents? Yes, and most are. Deloitte’s 2026 survey of organizations already piloting agents found 5 percent with processes highly prepared for them. The AI assessment is necessary; it does not answer the agentic question.

We already had an AI maturity assessment. Do we need another one? Not a repeat of it. The AI assessment is the base and should be reused. What is missing is a screen of each candidate workflow against the five questions that only matter once something acts, which most AI assessments never asked.

What is the first thing to check? Whether the workflow needs to act at all. If it does: who is accountable for what it does, under what identity, and whether anyone could reconstruct afterwards why it did it.

Sources (verified September 2026)

Weighing build versus buy?

Tell us about the workflow and the constraints. We will tell you honestly whether a fit-first build is the right call.

Talk to us